Expose explicit fused attention selection in the C API

[Gajesh2007]

Summary

Expose the existing C++ force_fused attention option through a new C entry point, mlx_fast_scaled_dot_product_attention_with_force_fused. Callers that must bound temporary allocation can request a fused kernel and receive an error for unsupported shapes instead of a composed score tensor.

The existing C entry point and ABI remain available and delegate with force_fused=false.

Before

flowchart LR
  A[C attention caller] --> B[mlx_fast_scaled_dot_product_attention]
  B --> C[C++ force_fused hardcoded false]
  C --> D[Default fused or composed routing]

Loading

After

flowchart LR
  A[Existing C caller] --> B[Existing entry point]
  B --> C[New entry point with false]
  N[Caller requiring fused allocation] --> T[New entry point with true]
  C --> D[Existing default routing]
  T --> E[C++ fused kernel or explicit refusal]

Loading

Validation

Five Swift/C API tests passed on M4 Max: explicit-false equality with the existing entry point, fused pitched views/masks/sinks against a dense reference, GQA8, D512 refusal, and a no-score-materialization allocation check. The wide-head cases use the companion MLX tile correction; the Swift package pins both dependencies together.

Used by the separately reserved prefill candidate in Layr-Labs/mlx-swift-lm#142 and Layr-Labs/d-inference#860.